Apple Intelligence Gets China Approval Running Alibaba's Qwen On-Device

China's Cyberspace Administration (CAC) approved Apple Intelligence on July 15, 2026, clearing it as one of seven approved on-device generative AI services for smartphones in China. The version of Apple Intelligence running in China does not use Apple's own models. It uses Alibaba's Qwen, compressed from 54GB to under 4GB by PrismML and running entirely on-device on iPhone 15 and newer hardware.

Alibaba's US-listed shares rose 4% on the news. The deal itself was struck in February 2025 — Apple and Alibaba announced it in public — but Beijing's content filtering requirements and the CAC's security evaluation process delayed the rollout for more than a year. The approval on July 15 ends that wait.

This is the first time a major Western technology company has shipped a Chinese AI model as the default intelligence layer on its consumer device for the Chinese market.

Why This Took 17 Months From Deal to Approval

Apple and Alibaba announced their partnership in February 2025. The deal got regulatory scrutiny in China for content moderation compliance, data handling requirements, and national security review before any Apple Intelligence feature could go live.

China requires generative AI services deployed to Chinese users to pass a CAC registration and evaluation process. The evaluation checks whether the AI's outputs comply with Chinese content regulations — which include restrictions on political content, historical interpretations, and a range of topics that are handled differently in Chinese AI outputs than in models trained primarily on Western internet data.

Qwen is a Chinese model trained by Alibaba with these requirements as design constraints rather than post-hoc filters. That is the reason Apple partnered with Alibaba rather than attempting to get Apple's own models approved: Alibaba's model was built to pass Chinese content review. Apple's models were not.

The CAC also evaluated on-device data handling — whether user data leaves the device, whether inference results are logged, and whether the system can be audited by Chinese regulators. On-device inference addresses some of these concerns: if the model runs locally and no query data is sent to a server, there is less data to regulate. PrismML's compression of Qwen to under 4GB is what made fully on-device inference viable on iPhone 15 hardware.

The Technical Work: How 27 Billion Parameters Fit on an iPhone

The standard Qwen model at 27 billion parameters weighs approximately 54GB. An iPhone 15 has 8GB of unified memory. Running a 54GB model on 8GB of RAM is not possible in standard deployment. PrismML, a Khosla Ventures-backed spinout from the California Institute of Technology, solved this through compression.

PrismML published compressed versions of Alibaba's open-source Qwen model, reducing it from roughly 54GB to under 4GB while retaining sufficient performance for the use cases Apple Intelligence requires: text generation, image understanding, writing assistance, and summarization.

The compression approach combines quantization (reducing the precision of model weights from 32-bit or 16-bit floating point to 4-bit or lower representations), pruning (removing weights that contribute minimally to output quality), and knowledge distillation (training a smaller model to replicate the behavior of the larger one). PrismML has not published the full technical details of their compression pipeline.

The result is a version of Qwen's 27B model that runs entirely on the Apple Neural Engine inside iPhone 15 and newer devices. No query is sent to a remote server. No Alibaba or Apple server sees the user's input. Inference is local. The output is generated on the device in a fraction of a second.

This matters for both privacy and performance. Privacy: local inference means no data leaves the device, which addresses both user concerns and regulatory requirements. Performance: network round-trip latency is eliminated, meaning Apple Intelligence responses in China are as fast as responses elsewhere — faster than a cloud-based model serving Chinese users from a data center would be.

What Apple Intelligence Features Work in China

The CAC approval covers Apple Intelligence experiences within iOS, iPadOS, macOS, and visionOS for Chinese users. The specific features vary by use case and continue to expand, but the core approved capabilities include:

Writing assistance (rewriting, proofreading, summarizing in Notes, Mail, and third-party apps that support the API), image understanding and generation within Photos and the system camera, Siri integration with Qwen-powered responses for general knowledge queries, and document summarization in Safari and Files.

The features that involve political content, sensitive historical topics, or information categories restricted under Chinese law are handled by Qwen's built-in content policies — the same policies that got the model through CAC review. Apple has not published a feature-by-feature comparison of what is available in China versus other markets.

Alibaba's Position After the Approval

The approval is a significant commercial win for Alibaba. Qwen runs on Apple hardware for Chinese users — a population of over one billion potential iPhone users. Every time an Apple Intelligence feature uses the Qwen model, Alibaba's technology is central to the experience.

The commercial arrangement between Apple and Alibaba has not been disclosed in full. The deal was announced as a partnership, not a licensing payment. Whether Alibaba receives per-inference revenue, a flat fee, or some other form of compensation for Qwen's deployment on Apple hardware is not public.

The 4% share price increase on July 15 reflects market confidence that the Apple deployment adds meaningful strategic value to Alibaba — either through direct revenue, brand association with premium consumer hardware, or data about how Qwen performs in real-world consumer deployment.

Our Analysis: What This Means for the AI Model Market in China

The Apple-Qwen approval sets a precedent. Apple is not the only Western tech company trying to deploy AI in China. Google, Microsoft, and Meta all have Chinese market considerations. The precedent Apple has now established is: to get AI features approved for Chinese users, use a Chinese model from a CAC-approved provider.

That narrows the realistic options. Alibaba Qwen and Baidu's ERNIE are the two most established CAC-compliant AI platforms. The report notes that Baidu is also involved in the Apple approval — possibly as a secondary model or for specific feature areas. Other Chinese model providers are in various stages of CAC registration.

For developers building apps targeting Chinese users: any generative AI feature requires either a CAC-registered model or no AI features in China. The CAC does not approve foreign models for Chinese AI features. Apple solved this by partnering with Alibaba. Independent developers need to make the same choice at their application layer.

The broader geopolitical implication is that China has effectively ensured Chinese AI models are the infrastructure layer for AI features on every major consumer device sold to Chinese users. Apple ships Qwen. Huawei ships its own models. Samsung ships Chinese models in China through agreements with local providers. The hardware is global. The AI layer is Chinese.

Key Takeaways

• CAC approved Apple Intelligence in China on July 15, 2026 — one of seven approved on-device generative AI services for smartphones
• Alibaba Qwen powers Apple Intelligence in China — not Apple's own models; deal was struck February 2025, delayed 17 months by CAC content review
• PrismML compressed Qwen from 54GB to under 4GB — allowing full on-device inference on iPhone 15 and newer with 8GB unified memory
• Alibaba US-listed shares +4% on approval news; Baidu also involved in the integration
• All inference is on-device: no user data sent to Alibaba or Apple servers for AI features in China
• Precedent set: Western tech companies deploying AI in China must use CAC-registered Chinese models; Qwen and ERNIE are the established options
• For developers: any app with generative AI features targeting Chinese users requires a CAC-registered model — foreign models are not approved for Chinese AI features